fix(sensors): floor the angles-to-columns map index in integer arithmetic

[janickm]

Summary

_init_angles_to_columns_map turns the nearest-neighbour flat element index (column-major, so flat = column * n_rows + row) into a column index by dividing by n_rows and truncating on the cast to the map dtype:

(idxs / self.n_rows).to(self.angles_to_columns_map_dtype)

The division runs in float32, which represents consecutive integers only below 2^24. Above that the quotient rounds up at every flat index just past a multiple of n_rows, moving those cells into the next column:

model	elements	wrong indices
128 x 3600 (today's Hesai)	460,800	0
512 x 14400 (largest map here)	7,372,800	0
128 x 140000	17,920,000	8,928
64 x 300000	19,200,000	37,856
128 x 200000	25,600,000	68,928

First divergence at flat index 16,777,343: float gives column 131073, exact floor gives 131072.

Not a numpy-version issue

Unlike the sibling fixes in #172 and #173, this is version-independent — the same counts reproduce under torch 2.7.0 and 2.13.0. It is latent headroom, not a live regression.

Today's maps stay well below the limit, so this is behaviour-preserving: the maps for both Hesai models come out bit-identical. It removes the dependency on that headroom and makes the line do what its comment already claimed ("dividing with the total number of rows and flooring").

torch.div's rounding_mode has been available since torch 1.8, well below the 1.12.1 pinned for the python 3.8 toolchain (verified against the actual pinned wheel).

Tests

A model with >2^24 elements would need a >16.7M-point KD-tree, far too expensive for a unit test. So the new test stubs the nearest-neighbour lookup to return flat indices straddling 2^24, which still exercises the real conversion inside _init_angles_to_columns_map. It fails on the unfixed code for all 16 parameterisations.

An earlier version of this test asserted the arithmetic directly rather than going through the model — it passed on the unfixed code, so it would not have caught a revert. The stub version is bound to the implementation.

A second test pins the mapped columns to the model's valid [0, n_columns - 1] range.

Validation

check	result
new test vs pre-fix	16 of 16 parameterisations fail
new test with fix	all pass
angles_to_columns_map vs pre-fix	bit-identical (0/7,372,800 and 0/2,457,600)
//ncore/impl/data:all, //ncore/impl/sensors:all, //tools/...	19 pass, python 3.8 (numpy 1.19.5, torch 1.12.1) and 3.11
numpy 1.26.4 + torch 2.7.0 / numpy 2.3.5 + torch 2.13.0	maps bit-identical, all 614,400 elements pass FOV asserts
bazel run format, ty	clean

Downstream

Ran NRE against this branch via --test_env=PYTHONPATH (sentinel-verified the override was live): //libs/vren:lidars_test, //nre/models/gaussians:gsplat_lidar_test, //nre/datasets/samplers:all — 8 pass.

Rebased onto main after #172 and #173 merged; re-validated there (19 bazel targets, both numpy/torch combinations, maps bit-identical to current main).